Make feedback resilient without LangFuse

Author: KostasCherv

cloudrun/service.yaml

@@ -115,3 +115,5 @@ spec:
             # langsmith-api-key / langfuse-* secrets when tracing matters.
             - name: LANGSMITH_TRACING
               value: "false"
+            - name: LANGFUSE_ENABLED
+              value: "false"

docs/env-vars-production.md

@@ -50,6 +50,7 @@ All secrets should be stored in **Google Secret Manager** and referenced via `va
 | `COHERE_API_KEY` | yes | Enables cross-encoder reranking |
 | `LANGFUSE_PUBLIC_KEY` | yes | |
 | `LANGFUSE_SECRET_KEY` | yes | |
+| `LANGFUSE_ENABLED` | no | Set to `true` only when both LangFuse keys are configured; otherwise feedback remains stored locally |
 | `LANGFUSE_ENV` | no | `prod` |
 | `LANGFUSE_RELEASE` | no | Git SHA or semver, e.g. `v1.2.3` |
 | `SENTRY_DSN` | yes | Error tracking; unset disables it entirely |

scripts/deploy.sh

@@ -53,7 +53,7 @@ gcloud container images add-tag "$TAG" "$LATEST" --quiet
 # ── Deploy ────────────────────────────────────────────────────────────────────
 
 # Inject the SHA-tagged image (and project id) so Cloud Run sees a spec change
-TMP_YAML=$(mktemp /tmp/service-XXXXXX.yaml)
+TMP_YAML=$(mktemp /tmp/service-XXXXXX)
 sed "s|gcr.io/PROJECT_ID/cortex:latest|$TAG|g" "$SERVICE_YAML" > "$TMP_YAML"
 
 if ! grep -q "$TAG" "$TMP_YAML"; then

src/api/routers/sessions.py

@@ -772,7 +772,7 @@ async def submit_run_feedback(
     body: RunFeedbackRequest,
     current_user: AuthenticatedUser = Depends(get_authenticated_user),
 ):
-    """Record simple user feedback for a completed run in LangFuse."""
+    """Record user feedback locally and mirror it to LangFuse when available."""
     session = await get_session(session_id, current_user.user_id)
     if session is None:
         raise HTTPException(
@@ -795,6 +795,15 @@ async def submit_run_feedback(
             status_code=409,
             detail="Feedback can only be submitted for completed runs.",
         )
+    comment: str | None = None
+    if body.comment is not None:
+        trimmed = " ".join(body.comment.strip().split())
+        if not trimmed:
+            raise HTTPException(status_code=400, detail="Feedback comment cannot be empty.")
+        if len(trimmed) > 500:
+            raise HTTPException(status_code=400, detail="Feedback comment is too long.")
+        comment = trimmed
+
     trace_id = run.langfuse_trace_id
     observation_id = run.langfuse_observation_id
     if not trace_id:
@@ -806,68 +815,48 @@ async def submit_run_feedback(
                 query=run.query,
                 report=run.report,
             )
-        except Exception as exc:
-            raise HTTPException(
-                status_code=502, detail=f"Could not link run to LangFuse: {exc}"
-            )
-        if not trace_id:
-            raise HTTPException(
-                status_code=502, detail="Could not link run to LangFuse."
-            )
-        linkage_updated = await update_session_run(
-            run_id=run_id,
-            user_id=current_user.user_id,
-            session_id=session_id,
-            patch={
+        except Exception:
+            # LangFuse is optional. A missing or unavailable observability backend
+            # must not prevent the product from accepting user feedback.
+            logger.warning("Could not create LangFuse feedback anchor", exc_info=True)
+            trace_id = None
+            observation_id = None
+
+    submitted_at = datetime.now(UTC).isoformat()
+    patch: dict[str, object] = {
+        "feedback_submitted_at": submitted_at,
+        "feedback_helpful": body.helpful,
+    }
+    if trace_id:
+        patch.update(
+            {
                 "langfuse_trace_id": trace_id,
                 "langfuse_observation_id": observation_id,
-            },
-        )
-        if not linkage_updated:
-            raise HTTPException(
-                status_code=404,
-                detail=f"Run '{run_id}' not found in session '{session_id}'.",
-            )
-
-    comment: str | None = None
-    if body.comment is not None:
-        trimmed = " ".join(body.comment.strip().split())
-        if not trimmed:
-            raise HTTPException(
-                status_code=400, detail="Feedback comment cannot be empty."
-            )
-        if len(trimmed) > 500:
-            raise HTTPException(status_code=400, detail="Feedback comment is too long.")
-        comment = trimmed
-
-    try:
-        submit_user_feedback_score(
-            trace_id=trace_id,
-            observation_id=observation_id,
-            helpful=body.helpful,
-            comment=comment,
-        )
-    except Exception as exc:
-        raise HTTPException(
-            status_code=502, detail=f"Could not submit LangFuse feedback: {exc}"
+            }
         )
-
-    submitted_at = datetime.now(UTC).isoformat()
     updated = await update_session_run(
         run_id=run_id,
         user_id=current_user.user_id,
         session_id=session_id,
-        patch={
-            "feedback_submitted_at": submitted_at,
-            "feedback_helpful": body.helpful,
-        },
+        patch=patch,
     )
     if not updated:
         raise HTTPException(
             status_code=404,
             detail=f"Run '{run_id}' not found in session '{session_id}'.",
         )
 
+    if trace_id:
+        try:
+            submit_user_feedback_score(
+                trace_id=trace_id,
+                observation_id=observation_id,
+                helpful=body.helpful,
+                comment=comment,
+            )
+        except Exception:
+            logger.warning("Could not mirror user feedback to LangFuse", exc_info=True)
+
     return {
         "session_id": session_id,
         "run_id": run_id,

tests/test_api.py

@@ -481,6 +481,71 @@ def test_submit_run_feedback_backfills_langfuse_linkage_when_missing():
     )
 
 
+def test_submit_run_feedback_succeeds_when_langfuse_is_not_configured():
+    mock_session = Session(
+        session_id="session-1",
+        runs=[SessionRun(run_id="run-1", query="q", source_urls=[], report="", created_at="2026")],
+        conversation=[],
+        created_at="2026",
+    )
+
+    with (
+        patch("src.api.routers.sessions.get_session", new=AsyncMock(return_value=mock_session)),
+        patch(
+            "src.api.routers.sessions.create_feedback_anchor_for_run",
+            side_effect=RuntimeError("LangFuse is not configured"),
+        ),
+        patch("src.api.routers.sessions.submit_user_feedback_score") as mock_score,
+        patch(
+            "src.api.routers.sessions.update_session_run", new=AsyncMock(return_value=True)
+        ) as mock_update,
+    ):
+        response = client.post(
+            "/sessions/session-1/runs/run-1/feedback",
+            json={"helpful": False},
+        )
+
+    assert response.status_code == 200
+    mock_score.assert_not_called()
+    assert mock_update.await_args.kwargs["patch"]["feedback_helpful"] is False
+
+
+def test_submit_run_feedback_succeeds_when_langfuse_score_fails():
+    mock_session = Session(
+        session_id="session-1",
+        runs=[
+            SessionRun(
+                run_id="run-1",
+                query="q",
+                source_urls=[],
+                report="",
+                created_at="2026",
+                langfuse_trace_id="trace-1",
+            )
+        ],
+        conversation=[],
+        created_at="2026",
+    )
+
+    with (
+        patch("src.api.routers.sessions.get_session", new=AsyncMock(return_value=mock_session)),
+        patch(
+            "src.api.routers.sessions.submit_user_feedback_score",
+            side_effect=RuntimeError("LangFuse unavailable"),
+        ),
+        patch(
+            "src.api.routers.sessions.update_session_run", new=AsyncMock(return_value=True)
+        ) as mock_update,
+    ):
+        response = client.post(
+            "/sessions/session-1/runs/run-1/feedback",
+            json={"helpful": True},
+        )
+
+    assert response.status_code == 200
+    assert mock_update.await_args.kwargs["patch"]["feedback_helpful"] is True
+
+
 def test_submit_run_feedback_rejects_duplicate_submission():
     mock_session = Session(
         session_id="session-1",